This is a repost of a number of messages I sent to the Facebook pages of JSesh in March 2024.

So Miyagawa has published a few experiments with the Claude large language model, with amazing results on translitteration (but really bad results on actual hieroglyphs). I have tried it too, and I must say I'm quite impressed. The system has produced perfectly good translations of passages from the LRL.

I gave it the sentence :

ḫr tw=k m-sšr m-bꜣḥ pꜣ (j)m(y)-r(ꜣ) mšꜥ.w mn btꜣ m-dꞽ=f ḫr bwpw rmṯ nb smꞽ=k m-bꜣḥ=f

it answered :

This text appears to be written in Ancient Egyptian hieroglyphs. Here's the transliteration of the hieroglyphs: ḫr tw=k m-sšr m-bꜣḥ pꜣ (j)m(y)-r(ꜣ) mšꜥ.w mn btꜣ m-dꞽ=f ḫr bwpw rmṯ nb smꞽ=k m-bꜣḥ=f Translation: *“Now, you are in good standing before the General. There is no fault with him. And no man has reported you before him.” This seems to be a statement about someone's good reputation in front of a military leader (General). The speaker assures the listener that they are in good standing with the General, there are no issues with the General himself, and no one has spoken negatively about the listener to the General. Please note that without more context and due to the complexity of the Ancient Egyptian language, this translation might not be entirely accurate, but it provides a general idea of the message conveyed in the hieroglyphic text.

I was somehow impressed. Attempts to feed Unicode Hieroglyphs to the system did not fare as well.

I tried a part of the Kanaïs text :

i:r-pA-n:t*y-nb-i-w-f-r-t:h-rd-D54-D21:V13-A1*B1:Z2-M17-G17-O34:N35:Z2-D37:D37-D21:V31:X1-Q1-X1:O1-D4:N35:I9

N35:G1-R8A-A40-R8-X1:D21-M17-M17-I12:Z2-V30:Z2-O7-A42-D21-M17-D21:Z4-A47-Y1:N35-D34:D40

Directly giving it to Claude gave:

This appears to be written in Egyptian hieroglyphs. Here is a transliteration of the text:

• ỉmn-rꜥ nsw nṯrw ỉw.f r ỉr ḥtp dỉ nsw n kꜣ.k m ẖrt-hrw (...)*

Which has nothing to do with the original text.

I fed the text to my Ramses-translit system, which gave the translitteration:

ꞽr pꜣ nty nb ꞽw =f r thꞽ.t rmṯ.w m =sn ꞽ.dꞽ r k.t s.t ꞽr.n =f nꜣ nṯr.w nṯr.wt nb.w ḥw.t =ꞽ r ꞽry n ꜥḥꜣ

and then I asked Claude : Please translate the Late Egyptian Text ꞽr pꜣ nty nb ꞽw=f r thꞽ.t rmṯ.w m=sn ꞽ.dꞽ r k.t s.t ꞽr.n=f nꜣ nṯr.w nṯr.wt nb.w ḥw.t=ꞽ r ꞽry n ꜥḥꜣ

His answer was:

The text you provided is written in Late Egyptian, which was used during the New Kingdom period (around 1300-700 BCE). Here's the transliteration and translation: Transliteration: (...)

Translation: “As for anyone who shall take people from them and put them in another place, all the gods and goddesses of his estate have cursed him to be subject to fighting.”

This passage appears to be a warning or a curse against anyone who would forcibly remove people from their rightful place and relocate them elsewhere. The text states that such an act would incur the wrath of all the gods and goddesses associated with the transgressor's estate, cursing them to face conflict or warfare as a consequence of their actions Please note that translations of ancient Egyptian texts may vary slightly depending on the interpreter and the specific context of the inscription.

Ramses-translit badly rendered ꞽr.n=f instead of ꞽr n=f, but Claude rectified it, although I would interpret it as a subjunctive instead of a perfective.

This is impressive, but there is still the possibility that the text and its translation are in the training data of Claude.

A similar attempt, with a cascade Ramses-translit + Claude, gave relatively bad results on the trial of the members and the stomach (P. Torino 58004).

A last attempt, to try something which was unlikely to be in the training database, used P. Berlin 3048, l. 3. (Jansen-Winkeln, Inschriften der Spätzeit II, p. 169).

I gave it the following translitteration, which closely follows the very late Egyptian orthography, and is quite misleading.

šsp n=k dbn 5 ḥḏ pꜣ pr-ḥḏ ḥry-š=f ꞽnk dꞽ.t sw n=k ꞽw=w 10 ḥsb.t 24 ꜣbd 1 šmw sw 11 ꞽw mn dꞽ.t=ꞽ md.t nb ꞽ.ḏd sw ꞽrm=k m-dr ꞽr mḏꜣ.t

I admit had to look in Malinine Choix de textes juridiques to find parallels and improve my own translation.

My own translation would be “(I) have received from you, 5 deben of silver from the temple of Heryshef ; it is I who will give them back to you as 10 (deben) in year 24, first month of Shemou, day 11, without me having any complain to say with you, as a document has been done.”

The speaker has borrowed 5 deben of silver at a 100% interest.

I'm a bit unsure about the m-dr ꞽr mḏꜣ.t, but it works, and it parallels the more usual formula ꞽw pꜣy sš m ḏr.t=k “as long as this writing is in your possession”.

The sentence is not really easy, as words are omitted, especially pronouns. Malinine, I, p. 18, n. 5, refers to Gardiner, Inscription of Mes p. 18-19 for šsp ...n with the meaning of to receive from. Also, Anne-Claude Honnay, Laurence Neven, in Ramses Online: reference [legacy/131/quote/212/227].

I have done a bit of "cherry picking", that is, given a few possible translations by the system, I have chosen the “best” one. This being said, most of them were quite close to the one I have chosen.

I asked CLAUDE :

Please translate the following Late Egyptian text from the third intermediate period: šsp n=k dbn 5 ḥḏ pꜣ pr-ḥḏ ḥry-š=f ꞽnk dꞽ.t sw n=k ꞽw=w 10 ḥsb.t 24 ꜣbd 1 šmw sw 11 ꞽw mn dꞽ.t=ꞽ md.t nb ꞽ.ḏd sw ꞽrm=k m-dr ꞽr mḏꜣ.t

Its answer :

This is a passage written in Late Egyptian, which was the language spoken in ancient Egypt from around 1300 to 700 BCE. > Here's a transliteration and translation: Transliteration: (...) Translation: Receive for yourself 5 debens of silver from the treasury which is upon it. I shall cause it to be given to you. They are 10. Year 24, first month of summer, day 11. There is no word which I have spoken with you that I could deny before a document.

Some notes (made by "CLAUDE", not by myself):

• "Deben" was a unit of measurement for weight, often used for metals like silver and gold.
• The "treasury" likely refers to a temple or royal treasury.
• The date is given according to the ancient Egyptian civil calendar, which had three seasons (Akhet/Inundation, Peret/Growing, Shemu/Harvest) of four months each.
• The last sentence seems to be a legal clause, perhaps indicating that the terms are fixed and not open to verbal alteration, only to what is written. Please note that without more context about the document this is from, this is a fairly literal translation. Some interpretive adjustments might be made with more background information.

Claude (at the time) did not fully understand the sentence, and did not “see” what was happening between the two parties. Neither did it understand that ḥry-š=f was Heryshef. But it did a relatively decent attempt.

Note that So Miyagawa has since created the Thot IA, which uses Claude and Retrieval Augmented Generation to work on Egyptian Texts.